Rosalind Picard: Affective Computing, Emotion, Privacy, and Health | Lex Fridman Podcast #24

The following is a conversation with Rosalind Picard. She's a professor at mit, Director of the Effective Computing Research Group at the MIT Media Lab, and co founder of two companies, Effectiva and Empatica. Over two decades ago, she launched the field of effective computing with her book of the same name. This book described the importance of emotion in artificial and natural intelligence, the vital role of emotional communication has to the relationship between people in general and human robot interaction. I really enjoy talking with Roz over so many topics, including emotion, ethics, privacy, wearable computing, and her recent research in epilepsy and even love and meaning. This conversation is part of the Artificial Intelligence podcast. If you enjoy it, subscribe on YouTube, iTunes, or simply connect with me on Twitter lexfridman spelled F R I D. And now here's my conversation with Rosalind.

Picard more than 20 years ago. You've coined the term effective computing and led a lot of research in this area since then. As I understand, the goal is to make the machine detect and interpret the emotional state of a human being and adapt the behavior of the machine based on the emotional state. So how is your understanding of the problem space defined by effective computing changed in the past 24 years? So it's the scope, the applications, the challenges, what's involved. How has that evolved over the years?

Yeah, actually originally when I defined the term affective computing, it was a bit broader than just recognizing and responding intelligently to human emotion, although those are probably the two pieces that we've worked on the hardest. The original concept also encompassed machines that would have mechanisms that functioned like human emotion does inside them. It would be any computing that relates to, arises from, or deliberately influences human emotion. So the human computer interaction part is the part that people tend to see. Like if I'm really ticked off at my computer and I'm scowling at it and I'm cursing at it, and it just keeps acting, smiling and happy like that little paperclip used to do. Dancing, winking, that kind of thing just makes you even more frustrated, right? And I thought that stupid thing needs to see my affect and if it's going to be intelligent, which Microsoft researchers had worked really hard on, it actually had some of the most sophisticated AI in it at the time. That thing's going to actually be smart. It needs to respond to me and you, and we can send it very different signals.

So by the way, just a quick interruption. The clippy, maybe it's in word 95 or 98. I don't remember when it was born, but many people do. You Find yourself with that reference that people recognize what you're talking about. Still to this point, the.

I don't expect the newest students to these days, but I've mentioned it to a lot of audiences, like, how many of you know this Clippy thing? And still the majority of people seem to know it.

So Clippy kind of looks at maybe natural language processing what you were typing and tries to help you complete. I think I don't even remember what Clippy was. Was. Except annoying.

Yeah, it tried. Some people actually liked it. I would hear those stories. You miss it?

Well, I miss the, the. The annoyance. They felt like there's a. There's an element was there, somebody was there and we're in it together. And they were annoying. It's like, it's like a puppy that just doesn't get it that keeps ripping up the college.

And in fact, they could have done it smarter like a puppy if they had done. Like if. When you yell, yelled at it or cursed at it, if it had put its little ears back and its tail down and shirked off, probably people would have wanted it back, right? But instead, when you yelled at it, what did it do? It smiled, it winked, it danced. Right. If somebody comes to my office and I yell at them, they start smiling, winking and dancing. I'm like, I never want to see you again. So Bill Gates got a standing ovation when he said it was going away because people were so ticked. It was so emotionally unintelligent. Right. It was intelligent about whether you were writing a letter, what kind of help you needed for that context. It was completely unintelligent about, hey, if you're annoying your customer, don't smile in their face when you do it. So that kind of mismatch was something the developers just didn't think about. And intelligence at the time was really all about math and language and chess and games, problems that could be pretty well defined. Social emotional interaction is much more complex than chess or Go or any of the games that people are trying to solve. And in order to understand that required skills that most people in computer science actually were lacking personally.

Well, let's talk about computer science. Have things gotten better since the work, since the message, since you've really launched a field with a lot of research work in this space. I still find as a person like yourself who's deeply passionate about human beings and yet eminent in computer science, there still seems to be a lack of, sorry to say, empathy in us computer scientists. Yeah, well, or hasn't gotten better.

Let's just Say there's a lot more variety among computer scientists these days. Computer scientists are a much more diverse group today than they were 25 years ago. And that's good. We need all kinds of people to become computer scientists so that computer science reflects more what society needs. And there's brilliance among every personality type. So it need not be limited to people who prefer computers to other people.

How hard do you think it is? How your view of how difficult it is to recognize emotion or to create a deeply emotionally intelligent interaction, has it gotten easier or harder as you've explored it further? And how far away are we from cracking this? The. If you think of the Turing Test, solving the intelligence, looking at the Turing Test for emotional intelligence.

I think it is as difficult as I thought it was going to be. I think my prediction of its difficulty is spot on. I think the time estimates are always hard because they're always a function of society's love and hate of a particular topic. If society gets excited and you get hundreds of, you get thousands of researchers working on, on it for a certain application, that application gets solved really quickly. The general intelligence, the lack, the, the computer's complete lack of ability to have awareness of what it's doing, the fact that it's not conscious, the fact that there's no signs of it becoming conscious, the fact that it doesn't read between the lines, those kinds of things that we have to teach it explicitly what other people pick up implicitly. We don't see that changing yet. There aren't breakthroughs yet that lead us to believe that that's going to go any faster, which means that it's still going to be kind of stuck with a lot of limitations where it's probably only going to do the right thing in very limited, narrow, pre specified context where we can prescribe pretty much what's going to happen there. So I don't see the. It's hard to predict the date because when people don't work on it, it's infinite. When everybody works on it, you get a nice piece of it well solved in a short amount of time. I actually think there's a more important issue right now than the difficulty of it. And that's causing some of us to put the brakes on a little bit. Usually we're all just like, step on the gas, let's go faster. This is causing us to pull back and put the brakes on. And that's the way that some of this technology is being used in places like China right now. And that worries me so deeply that it's causing me to pull back myself on a lot of the things that we could be doing and, and try to get the community to think a little bit more about. Okay, if we're going to go forward with that, how can we do it in a way that puts in place safeguards that protects people?

So the technology we're referring to is just when a computer senses the human being, like the human face.

Yeah, right, yeah.

And so what? There's a lot of exciting things there, like forming a deep connection with a human being. But so what are your worries, how that could go wrong? Is it in terms of privacy? Is it in terms of other kinds of morality?

Let's dig into privacy. So here in the US if I'm watching a video of, say, a political leader, and in the US we're quite free, as we all know, to even criticize the President of the United States right here. That's not a shocking thing. It happens about every five seconds. But, but in China, what happens if you criticize the leader of the government? Right. And so people are very careful not to do that. However, what happens if you're simply watching a video and you make a facial expression that shows a little bit of skepticism? Right. Well, in, you know, here we're completely free to do that. In fact, we're free to fly off the handle and say anything we want usually. I mean, there are some restrictions when the athlete does this as part of the national broadcast. Maybe the teams get a little unhappy about picking that forum to do it. Right. But that's more a question of judgment. We have these freedoms and in places that don't have those freedoms, what if our technology can read your underlying affective state? What if our technology can read it even non contact, what if our technology can read it without your prior consent? And here in the US in my first company, we started Affectiva, we have worked super hard to turn away money and opportunities that try to read people's affect without their prior informed consent. And even the software that is licensable, you have to sign things saying you will only use it in certain ways, which essentially is get people's buy in. Right. Don't, don't do this without people agreeing to it. There are other countries where they're not interested in people's buy in, they're just going to use it, they're going to inflict it on you. And if you don't like it, you better not scowl in the direction of any sensors.

So one, let me just comment on a small tangent. Do you know with the idea of adversarial examples and deep fakes and so on, what you bring up is actually, in one sense, deep fakes provide a comforting protection that you can no longer really trust that the video of your face was legitimate. And therefore you always have an escape clause. If a government is trying, if a stable, balanced, ethical government is trying to accuse you of something, at least you have protection. You can say it was fake news, as is a popular term now.

Yeah, that's the general thinking of it. We know how to go into the video and see, for example, your heart rate and respiration and whether or not they've been tampered with. And we also can put like fake heart rate and respiration in your video now too. We decided we needed to do that after we, after we developed a way to extract it, we decided we also needed a way to jam it. And so the fact that we took time to do that other step too. Right. That was time that I wasn't spending making the machine more effectively intelligent. And there's a choice in how we spend our time, which is now being swayed a little bit less by this goal and a little bit more like by concern about what's happening in society and what kind of future do we want to build. And as we step back and say, okay, we don't just build AI to build AI to make Elon Musk more money or to make Amazon Jeff Bezos more money, good gosh, that's the wrong ethic. Why are we building it? What is the point of building AI? It used to be it was driven by researchers in academia to get papers published and to make a career for themselves and to do something cool because maybe it could be done. Now we realize that this is enabling rich people to get vastly richer. The poor, the divide is even larger. And is that the kind of future that we want? Maybe we want to think about. Maybe we want to rethink AI, maybe we want to rethink the problems in society that are causing the greatest inequity and rethink how to build AI that's not about a general intelligence, but that's about extending the intelligence and capability of the have nots so that we close these gaps in society.

Do you hope that kind of stuff stepping on the brake happens organically? Because I think still majority of the force behind AI is the desire to publish papers is to make money without thinking about the why do you hope it happens organically? Is there room for regulation?

Yeah, yeah, yeah, great questions. I prefer the, you know, they talk about the Carrot versus the stick. I definitely prefer the carrot to the stick. And you know, in our free world there's only so much stick, right? You're going to find a way around it. I generally think less regulation is better. That said, even though my position is classically carrot, no stick, no regulation, I think we do need some regulations in this space. I do think we need regulations around protecting people with their data, that you own your data, not Amazon, not Google. I would like to see people own their own data. I would also like to see the regulations that we have right now around lie detection, detection being extended to emotion recognition in general, that right now you can't use a lie detector on an employee when you're on a candidate, when you're interviewing them for a job. I think similarly we need to put in place protection around reading people's emotions without their consent and in certain cases, like characterizing them for a job and other opportunities. So I also think that when we're reading emotion that's predictive around mental health, that that should, even though it's not medical data, that that should get the kinds of protections that our medical data gets. What most people don't know yet is right now with your smartphone use and if you're wearing a sensor and you want to learn about your stress and your sleep and your physical activity and how much you're using your phone and your social interaction, all of that non medical data, when we put it together with machine learning now called AI, even though the founders of AI wouldn't have called it that, that capability can not only tell that you're calm right now or that you're getting a little stressed, but it can also predict how you're likely to be tomorrow, if you're likely to be sick or healthy, happy or sad, stressed or calm, especially when you're.

Tracking data over time, especially when we're.

Tracking a week of your data or more.

Do you have an optimism towards, you know, a lot of people on our phones are worried about this camera that's looking at us for the most part, on balance. Do you, Are you optimistic about the benefits that can be brought from that camera that's looking at billions of us, or should we be more worried?

I think we should be a little bit more worried about who's looking at us and listening to us, the device sitting on your countertop in your kitchen, whether it's, you know, Alexa or Google Home or Apple Siri, these devices want to listen while they say, ostensibly to help us. And I think there are great people in these Companies who do want to help people. Let me not brand them all bad. I'm a user of products from all, all of these companies. I'm naming all the A companies. Alphabet, Apple, Amazon. They are awfully big companies, right? They have incredible power. And you know, what if, what if China were to buy them, right? And suddenly all of that data were not part of Free America, but all of that data were part of somebody who just wants to take over the world and you submit to them. And guess what happens if you so much as smirk the wrong way when they say something that you don't like? Well, they have re education camps, right? That's a nice word for them, by the way. They have a surplus of organs for people who have surgery these days. They don't have an organ donation problem because they take your blood and they know you're a match. And the doctors are on record of taking organs from people who are perfectly healthy and not prisoners. They're just simply not the favored ones of the government. And you know, that's a pretty freaky evil society. And we can use the word evil there.

I was born in the Soviet Union. I can certainly connect to the worry that you're expressing at the same time. Probably both you and I and you very much so. You know, there's an exciting possibility that you can have a deep connection with the machine.

Yeah, yeah, right. So those of us I've admitted students who say that they, you know, when you list, like, who do you most wish you could have lunch with or dinner with, right? And they'll write like, I don't like people, I just like computers. And one of them said to me once when I had this party at my house, I want you to know this is my only social event of the year. My one social event of the year, okay? Now this is a brilliant machine learning person, right? And we need that kind of brilliance in machine learning. And I love that computer science welcomes people who love people and people who are very awkward around people. I love that this is a field that anybody could join. We need all kinds of people, people. And you don't need to be a social person. I'm not trying to force people who don't like people to suddenly become social. At the same time, if most of the people building the AIs of the future are the kind of people who don't like people, we've got a little bit of a problem.

Hold on a second. So let me, let me push back on that. So don't you think a large percentage of the world can. You know, there's loneliness.

There is a huge problem with loneliness.

That'S growing, and so there's a longing for connection. Do you.

If you're lonely, you're part of a big and growing group.

Yes. So we're in it together. I guess if you're lonely, you're not alone. That's a good line. But do you think there's. You talked about some worry, but do you think there's an exciting possibility that something like Alexa and these kinds of tools can alleviate that loneliness in a way that other humans can't?

Yeah. Yeah, definitely. I mean, a great book can kind of alleviate loneliness because you just get sucked into this amazing story and you can't wait to go spend time with that character. Right. And they're not a human character. There is a human behind it. But. But yeah, it can be an incredibly delightful way to pass the hours, and it can meet needs, even. You know, I don't read those trashy romance books, but somebody does, right. And what are they getting from this? Well, probably some of that feeling of being there. Right. Being there in that social moment, that romantic moment, or connecting with somebody. I've had a similar experience reading some science fiction books. Right. And connecting with the character. Orson Scott Card, you know, just amazing writing. And Ender's Game and Speaker for the Dead. Terrible title. But those kind of books that pull you into a character and you feel like you're. You feel very social. It's very connected, even though it's not responding to you. And a computer, of course, can respond to you, so it can deepen it. Right? You can have a very deep connection. Much more than the movie her, Right. You know, plays up. Right.

Well, much more. I mean, movie her is already a pretty. Pretty deep connection.

Right?

Well, but it. But it's just a movie, right? It's scripted. It's just, you know, but I. I mean, like, there can be a real interaction where the character can learn and you can learn. You could imagine it not just being you and. And one character. You can imagine a group of characters. You can imagine a group of people and characters, human and AI connecting, where maybe a few people can't can't sort of be friends with everybody, but the few people and their AIs can befriend more people. There can be an extended human intelligence in there where each human can connect with more people that way. But it's. It's still very limited. But there are just. What I mean is there are many more possibilities than what's in that movie.

So there's a tension here. So one, you express a really serious concern about privacy, about how governments can misuse the information and there's the possibility of this connection. So let's look at Alexa. So personal assistants, for the most part, as far as I'm aware, they ignore your emotion. They ignore even the context or the existence of you. The intricate, beautiful, complex aspects of who you are. Except maybe aspects of your voice that help it recognize for speech recognition. Do you think they should move towards trying to understand your emotion?

All of these companies are very interested in understanding human emotion. More people are telling Siri every day they want to kill themselves. Apple wants to know the difference between if a person is really suicidal versus if a person is just kind of fooling around with Siri. Right. The words may be the same. The tone of voice and what surrounds those words is pivotal to understand if they should respond in a very serious way, bring help to that person, or if they should kind of jokingly tease back, you know, ah, you just want to, you know, sell me for something else. Right? Like, like how do you respond when somebody says that? Well, you know, you do want to err on the side of being careful and taking it seriously. People want to know if the person is happy or stressed in part. Well, so let me give you an altruistic reason and a business profit motivated reason. And there are people in companies that operate on both principles. The altruistic people really care about their customers and really care about helping you feel a little better at the end of the day. And it would just make those people happy if they knew that they made your life better if you came home stressed and after talking with their product, you felt better. There are other people who maybe have studied the way affect affects decision making and prices. People pay and they know. I don't know if I should tell you, like the work of Jen Lerner on heartstrings and purse strings, you know, if we manipulate you into a slightly sadder mood, you'll pay more. Yeah, right. You'll pay more to change your situation. You'll pay more for something you don't even need to make yourself feel better. So, you know, if they sound a little sad, maybe I don't want to cheer them up. Maybe first I want to help them get something, a little shopping therapy. Right.

That helps them, which is really difficult for a company that's primarily funded on advertisements. So they're encouraged to, to get you.

To offer you products. Or Amazon that's primarily funded on you buying things from their Store. So I think we should be, you know, maybe we need regulation in the future to put a little bit of a wall between these agents that have access to our emotion and agents that want to sell us stuff. Maybe there needs to be a little bit more of a firewall in between those.

So maybe digging in a little bit on the interaction with Alexa. You mentioned, of course, a really serious concern about like, recognizing emotion if somebody is speaking of suicide or depression and so on. But what about the actual interaction itself? Do you think? So if I, if I, you know, you mentioned Clippy and being annoying. What is the objective function we're trying to optimize? Is it minimize annoyingness or minimize or maximize happiness or. But if we look at human, Human relations, I think that push and pull, the tension, the dance, you know, the annoying, the flaws, that's what makes it fun. So is there.

Yeah, yeah.

Is there a room for like, what. What is the upside down.

When you want to, you want to have a little push and pull. Think of kids sparring. Right. You know, I see my sons and they, one of them wants to provoke the other to be upset. And that's fun. And it's actually healthy to, to learn where your limits are, to learn how to self regulate. You can imagine a game where it's trying to make you mad and you're trying to show self control. And so if we're doing a AI Human interaction that's helping build resilience and self control, whether it's to learn how to not be a bully or how to turn the other cheek or how to deal with an abusive person in your life, then you might need an AI that pushes your buttons. Right. But in general, do you want an AI that pushes your buttons? Probably depends on your personality. I don't. I want one that's respectful, that is there to serve me and that is there to extend my ability to do things. I'm not looking for a rival. I'm looking for a helper. And that's the kind of AI I.

Put my money on your census for the majority of people in the world in order to have a rich experience. That's what they're looking for as well. So they're not. Look, if you look at the movie her. Spoiler alert. I, I believe the, the program, the, the woman in the movie, her leaves the yeah. Person for somebody else.

Right, Right.

Says they don't want to be dating anymore. Right. Like, do you. Your sense is if Alexis said, you know what, I actually had enough of you for a while. So I'M going to shut myself off. You don't see that as I'd say.

You'Re trash because I paid for you. Right? We've got to remember, and this is where this blending human AI as if we're equals is really deceptive because AI is something at the end of the day that my students and I are making in the lab. And we're choosing what it's allowed to say, when it's allowed to speak, what it's allowed to listen to, what it's allowed to act on, given the inputs that we choose to expose it to, what outputs it's allowed to have. It's all something made by a human. And if we want to make something that makes our lives miserable, fine. I, I wouldn't invest in it as a business, you know, unless it's just there for self regulation training. But I think we, you know, we need to think about what kind of future we want and actually your question, I really like the. What is the objective function? Is it to calm people down sometimes? Is it to always make people happy and calm them down? Well, there was a book about that, right? The Brave New World, you know, make everybody happy. Take your soma. If you're unhappy, take your happy pill. And if you refuse to take your happy pill, well, we'll, we'll threaten you by sending you to Iceland to live there. I lived in Iceland three years. It's a great place. Don't take your soma then go to Iceland. A little TV commercial there. I was a child there for a few years. It's a wonderful place. So that part of the book never scared me, but, but really, like, do we want AI to manipulate us into submission, into making us happy? Well, if you are a, you know, like a power obsessed, sick dictator individual who only wants to control other people to get your jollies in life, then yeah, you want to use AI to extend your power and your scale to force people into submission. If you believe that the human race is better off being given freedom and the opportunity to do things that might surprise you, then you want to use AI to extend people's ability. You want to build AI that extends human intelligence, that empowers the weak and helps balance the power between the weak and the strong. Not that gives more power to the strong.

So in this process of empowering people and sensing people, and what is your sense on emotion in terms of recognizing emotion, the difference between emotion that is shown and emotion that is felt? So yeah, emotion that is expressed on the surface through your face, your body and Various other things and what's actually going on deep inside on the biological level, on the neuroscience level or some kind of cognitive level.

Yeah, yeah. Whoa, no easy questions here.

Well, yeah, I'm sure there's no definitive answer, but what's your sense? How far can we get by just looking at the face?

We're very limited when we just look at the face, but we can get further than most people think. We can get. People think, hey, I have a great poker face, therefore all you're ever going to get from me is neutral. Well, that's naive. We can read with the ordinary camera on your laptop or on your phone. We can read from a neutral face. If your heart is racing, we can read from a neutral face. If your breathing is becoming irregular and showing signs of stress, we can read under some conditions that maybe I won't give you details on how your heart rate variability power is changing. That could be a sign of stress even when your heart rate is not necessarily accelerating.

So sorry. From physio senses or from the face.

From the color changes that you cannot even see, but the camera can see.

That's amazing. So you can get a lot of signal, but.

So we get things people can't see using a regular camera and from that we can tell things about your stress. So if you are just sitting there with a blank face thinking nobody can read my emotion. Well, you're wrong.

Right? So that's really interesting. But that's from sort of visual information from the face that's almost like cheating your way to the physiological state of the body by being very clever with.

What you can do with signal processing.

Signal processing. So that's really impressive. But if you just look at the stuff we humans can see, the smile, the smirks, the subtle, all the facial actions.

So then you can hide that on your face for a limited amount of time. Now if you're just going in for a brief interview and you're hiding it, that's pretty easy for most people. If you are, however, surveilled constantly everywhere you go, then it's going to say, gee, Lex used to smile a lot and now I'm not seeing so many smiles. And Roz used to, you know, laugh a lot and smile a lot very spontaneously. And now I'm only seeing these not so spontaneous looking smiles and only when she's asked these questions. You know, that's something.

Sleep.

We could look at that too. So now I have to be a little careful too when I say we. You think we can't read your emotion and we can it's not that binary. What we're reading is more some physiological changes that relate to your activation. Now that doesn't mean that we know everything about how you feel. In fact we still know very little about how you feel. Your thoughts are still private, your nuanced feelings are still completely private. We can't read any of that. So there's some relief that we can't read that. Even brain imaging can't read that. Wearables can't read that. However, as we read your body state changes and we know what's going on in your environment and we look at patterns of those over time, we can start to make some inferences about what you might be feeling. And that is where it's not just the momentary feeling but it's more your state stance toward things. And that could actually be a little bit more scary with certain kinds of governmental control freak people who want to know more about are you on their team or are you not and getting.

That information through over time. So you're saying there's a lot of signal by looking at the change over time.

Yeah.

So you've done a lot of exciting work both in computer vision and physiological sounds like wearables. What do you think is the best modality for what's the, the, the best window into the emotional soul? Like is it the face, is it the voice?

Depends what you want to know. It depends what you want to know. It depends what you want to know. Everything is informative. Everything we do is informative.

So for health and well being and things like that, do you find the wearable physio technos measuring physiological signals is the best for health based stuff.

So here I'm going to answer empirically with data and studies. We've been doing, we've been doing studies now these are currently running with lots of different kinds of people. But where we've published data and I can speak publicly to it, the data are limited right now to New England college students. So that's a small group among New England college students when they are wearing a wearable like the empatic embrace here that's measuring skin conductance, movement, temperature and when they are using a smartphone that is collecting their time of day of when they're texting, who they're texting, their movement around it, their gps, the weather information based upon their location and when it's using machine learning and putting all of that together and looking not just at right now but looking at your rhythm of behaviors over about a week, when we look at that we are very accurate at forecasting tomorrow's stress mood and happy, sad mood and health. And when we look at which pieces of that are most useful, first of all, if you have all the pieces, you get the best results. If you have only the wearable, you get the next best results. And that's still better than 80% accurate at forecasting tomorrow's levels.

Isn't that exciting? Because the wearable stuff with physiological information, it feels like it violates privacy less than the non contact face based methods.

Yeah, it's interesting. I think what people sometimes don't. You know, it's funny, in the early days people would say, oh, wearing something or giving blood is invasive, right? Whereas a camera is less invasive because it's not touching you. I think on the contrary, the things that are not touching you are maybe the scariest because you don't know when they're on or off and you don't know when and you don't know who's behind it. Right. A wearable, depending upon what's happening to the data on it, if it's just stored locally or if it's streaming and what it is being attached to, in a sense you have the most control over it because it's also very easy.

To just take it off.

Take it off. Right now it's not sensing me. So if I'm uncomfortable with what it's sensing now, I'm free, right? If I'm comfortable with what it's sensing then, and I happen to know everything about this one and what it's doing with it, so I'm quite comfortable with it, then I have control. I'm comfortable. Control is one of the biggest factors.